Reflection Releases Beam's Weights Under Apache 2.0 -- a 501B Open Model Pitched on Fewer Tokens per Task, Not Cheaper Tokens

On October 16, 2026, Reflection AI wrote: "Today, we have released Beam's model weights on Hugging Face under an Apache 2.0 license." Beam is "a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active, built for coding, reasoning, and agentic workloads." When Reflection announced it on October 5 the weights were not yet downloadable. They now ship in BF16, FP8 and NVFP4, with a speculative decoding drafter model.

What is in the release

• Architecture. Hugging Face's launch post describes a text-only model with a context window of 1M tokens and 52 layers: 43 use sliding-window attention over 512 tokens and 9 attend to the full context. Only 8 of 256 experts run per token, but "All 256 experts stay in memory", and the BF16 checkpoint is about 1 TB.
• Hardware. Reflection told VentureBeat the 8-bit and 4-bit versions fit on one eight-GPU H100, H200 or GB300 server. Full precision needs "tens of GPUs."
• Software. Day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl and NVIDIA Dynamo. Reflection also open-sourced Mirror CLI, a Rust-and-Python coding agent it uses internally, and says it launched with "an ecosystem of over 40 partners and integrations".
• Safety. A Beam Safety Technical Report was published alongside, and VentureBeat reports that Gray Swan, Irregular and SecureBio tested the model before release.

The pitch is cost per task

Reflection is not claiming Beam is the strongest open model. In its own coding table, VentureBeat notes, DeepSeek V4.1 Flash scores 90.6% on Terminal-Bench 2.1 against Beam's 80.1%. The claim is efficiency. Reflection says it penalised the model during reinforcement learning whenever it used unnecessary tokens. By VentureBeat's reading of Reflection's chart, Beam scores about 80% on Terminal-Bench 2.1 with roughly 25,000 tokens per attempt, while GLM-5.2 scores slightly lower with roughly 80,000.

CTO Ioannis Antonoglou explained the hard part: a naive penalty produces a model that "will just learn to not solve the task," because short answers are easier than solved ones. Reflection's answer was scale: more than 100 million rollouts over four weeks on 10,500 GB300 GPUs, and about 1.3 billion sandboxes in total across training, evaluation and ablations.

VentureBeat is careful about what that shows. Reflection's compute estimate "is an estimate, not a measurement", and it leaves out the cost of reading the prompt, which dominates long agent sessions. Asked for a dollar cost per finished coding task, VP of Product Joe Spisak said: "We don't really have anything concrete to share there." He also conceded that rivals "may have better kernels or fused kernels".

Provenance as a selling point

Reflection told VentureBeat it trained Beam from scratch and "does not train on any other model's outputs." Antonoglou argued that companies running agents cannot be sure a model was not "trained to just activate certain behaviors in certain scenarios", while adding: "This is not to say that Chinese models are bad." VentureBeat's own September survey suggests the question is live. Of 32 respondents running open-weight models, 56% could not say with confidence where their models originate, or did not know whether any came from a Chinese lab.

What remains uncertain. There is no published price yet, and no independent cost-per-task measurement. VentureBeat says it is running one across Beam, Muse Spark Open, DeepSeek V4.1 Flash and GLM-5.2. The FT's October 10 report that Nvidia is in early talks over Reflection is unresolved: Antonoglou called it "baseless rumors" but did not say whether talks are taking place.

Analysis: Two US open releases in two days split the buying decision cleanly. Meta's Muse Spark Open competes on capability, with a license that needs a separate deal from large inference resellers. Beam competes on tokens per finished task, under one of the most permissive open licenses. For teams that run coding agents on their own hardware, the second pitch is easier to test. Run your own task set through both and count tokens, not list prices. The open question is whether the token savings survive long prompts, which Reflection's own estimate leaves out.